Multi-frequency feature learning for building change detection using DFNet on S2Looking dataset

This study presents a novel DFNet-based deep learning framework for Building Change Detection (BCD) in high-resolution remote sensing imagery, utilizing the S2Looking dataset to address challenges such as sensor noise, light variations, and pseudo-changes. We propose a hybrid architecture integrating Dual-Frequency Feature Extraction (DFFE), Feature Enhancement with Variation Constraints (FEVC), and Multi-Frequency Domain Dense Interaction (MFDDI) modules to capture high- and low-frequency features, enhance feature discriminability, and fuse multi-scale information for robust change detection between pre- and post-change satellite images. The pipeline preprocesses images with Gaussian blur for noise reduction and applies data augmentation to improve generalization. It effectively captures high- and low-frequency features while enhancing feature representation through advanced attention mechanisms, leading to precise change detection. Evaluated on the S2Looking dataset, DFNet outperforms eight baselines, achieving state-of-the-art performance with an accuracy of 94.71%. Ablation studies confirm the efficacy of DFFE, FEVC, and MFDDI, with combined modules yielding 91.26% overall accuracy.